Implement llvm.x86.subborrow.32 and llvm.x86.subborrow.64

This commit is contained in:
Eduardo Sánchez Muñoz 2023-09-22 16:27:27 +02:00
parent 6ec63ed251
commit a617b89996
2 changed files with 68 additions and 0 deletions

View File

@ -51,6 +51,32 @@ pub(super) trait EvalContextExt<'mir, 'tcx: 'mir>:
this.write_scalar(Scalar::from_u8(c_out.into()), &this.project_field(dest, 0)?)?;
this.write_immediate(*sum, &this.project_field(dest, 1)?)?;
}
// Used to implement the `_subborrow_u32` and `_subborrow_u64` functions.
// Computes a - b with input and output borrow. The input borrow is an 8-bit
// value, which is interpreted as 1 if it is non-zero. The output borrow is
// an 8-bit value that will be 0 or 1.
// https://www.intel.com/content/www/us/en/docs/cpp-compiler/developer-guide-reference/2021-8/subborrow-u32-subborrow-u64.html
"subborrow.32" | "subborrow.64" => {
if unprefixed_name == "subborrow.64" && this.tcx.sess.target.arch != "x86_64" {
return Ok(EmulateByNameResult::NotSupported);
}
let [b_in, a, b] = this.check_shim(abi, Abi::Unadjusted, link_name, args)?;
let b_in = this.read_scalar(b_in)?.to_u8()? != 0;
let a = this.read_immediate(a)?;
let b = this.read_immediate(b)?;
let (sub, overflow1) = this.overflowing_binary_op(mir::BinOp::Sub, &a, &b)?;
let (sub, overflow2) = this.overflowing_binary_op(
mir::BinOp::Sub,
&sub,
&ImmTy::from_uint(b_in, a.layout),
)?;
let b_out = overflow1 | overflow2;
this.write_scalar(Scalar::from_u8(b_out.into()), &this.project_field(dest, 0)?)?;
this.write_immediate(*sub, &this.project_field(dest, 1)?)?;
}
name if name.starts_with("sse.") => {
return sse::EvalContextExt::emulate_x86_sse_intrinsic(

View File

@ -13,6 +13,14 @@ mod x86 {
(c_out, sum)
}
fn sbb(b_in: u8, a: u32, b: u32) -> (u8, u32) {
let mut sum = 0;
// SAFETY: There are no safety requirements for calling `_subborrow_u32`.
// It's just unsafe for API consistency with other intrinsics.
let b_out = unsafe { arch::_subborrow_u32(b_in, a, b, &mut sum) };
(b_out, sum)
}
pub fn main() {
assert_eq!(adc(0, 1, 1), (0, 2));
assert_eq!(adc(1, 1, 1), (0, 3));
@ -22,6 +30,19 @@ mod x86 {
assert_eq!(adc(1, u32::MAX, u32::MAX), (1, u32::MAX));
assert_eq!(adc(2, u32::MAX, u32::MAX), (1, u32::MAX));
assert_eq!(adc(u8::MAX, u32::MAX, u32::MAX), (1, u32::MAX));
assert_eq!(sbb(0, 1, 1), (0, 0));
assert_eq!(sbb(1, 1, 1), (1, u32::MAX));
assert_eq!(sbb(2, 1, 1), (1, u32::MAX)); // any non-zero borrow acts as 1!
assert_eq!(sbb(u8::MAX, 1, 1), (1, u32::MAX));
assert_eq!(sbb(0, 2, 1), (0, 1));
assert_eq!(sbb(1, 2, 1), (0, 0));
assert_eq!(sbb(2, 2, 1), (0, 0));
assert_eq!(sbb(u8::MAX, 2, 1), (0, 0));
assert_eq!(sbb(0, 1, 2), (1, u32::MAX));
assert_eq!(sbb(1, 1, 2), (1, u32::MAX - 1));
assert_eq!(sbb(2, 1, 2), (1, u32::MAX - 1));
assert_eq!(sbb(u8::MAX, 1, 2), (1, u32::MAX - 1));
}
}
@ -37,6 +58,14 @@ mod x86_64 {
(c_out, sum)
}
fn sbb(b_in: u8, a: u64, b: u64) -> (u8, u64) {
let mut sum = 0;
// SAFETY: There are no safety requirements for calling `_subborrow_u64`.
// It's just unsafe for API consistency with other intrinsics.
let b_out = unsafe { arch::_subborrow_u64(b_in, a, b, &mut sum) };
(b_out, sum)
}
pub fn main() {
assert_eq!(adc(0, 1, 1), (0, 2));
assert_eq!(adc(1, 1, 1), (0, 3));
@ -46,6 +75,19 @@ mod x86_64 {
assert_eq!(adc(1, u64::MAX, u64::MAX), (1, u64::MAX));
assert_eq!(adc(2, u64::MAX, u64::MAX), (1, u64::MAX));
assert_eq!(adc(u8::MAX, u64::MAX, u64::MAX), (1, u64::MAX));
assert_eq!(sbb(0, 1, 1), (0, 0));
assert_eq!(sbb(1, 1, 1), (1, u64::MAX));
assert_eq!(sbb(2, 1, 1), (1, u64::MAX)); // any non-zero borrow acts as 1!
assert_eq!(sbb(u8::MAX, 1, 1), (1, u64::MAX));
assert_eq!(sbb(0, 2, 1), (0, 1));
assert_eq!(sbb(1, 2, 1), (0, 0));
assert_eq!(sbb(2, 2, 1), (0, 0));
assert_eq!(sbb(u8::MAX, 2, 1), (0, 0));
assert_eq!(sbb(0, 1, 2), (1, u64::MAX));
assert_eq!(sbb(1, 1, 2), (1, u64::MAX - 1));
assert_eq!(sbb(2, 1, 2), (1, u64::MAX - 1));
assert_eq!(sbb(u8::MAX, 1, 2), (1, u64::MAX - 1));
}
}